Presentation is loading. Please wait.

Presentation is loading. Please wait.

Tim Hewison (EUMETSAT) (GRWG Chair)

Similar presentations


Presentation on theme: "Tim Hewison (EUMETSAT) (GRWG Chair)"— Presentation transcript:

1 Tim Hewison (EUMETSAT) (GRWG Chair)
Special Issue of the IEEE TGRS on “Inter-Calibration of Satellite Instruments”: Special Issue of the IEEE TGRS on “Inter-Calibration of Satellite Instruments”: 12 November 2018 Defining Primary GSICS References - an Evolving Concept for Inter-Calibration Products Tim Hewison (EUMETSAT) (GRWG Chair)

2 Background Motivation and Overall Proposal

3 Current Situation GSICS Inter-Calibration Products
Defined by (Monitoring – Reference) Instruments GEO-LEO IR Products based on MetopA/IASI Based on SNO method Now in Pre-Operational mode + Aqua/AIRS as transfer to characterise diurnal variations Developing Delta Corrections for MetopB/IASI GEO-LEO VIS Products based on Aqua/MODIS Based on DCC method Soon to be in Demonstration mode

4 Motivation to Change New reference instruments becoming available
MetopB/IASI, SNPP/CrIS, ... VIIRS, ... Need to define Archive Recalibration Corrections To support generation of FCDRs Aim to blend multiple references After bias adjustment Define a Primary GSICS Reference by consensus More methods becoming available DCC, Lunar, Rayleigh, ... Ultimately aim to combine multiple methods

5 Requirements Long-term continuity of GSICS Products
Without calibration jumps between references Although their uncertainty will change... Ensuring Traceability back to single Primary Reference For each Spectral Band Radiometric Consistency Cover a broad range of applicability

6 Proposed Solution – Define Primary GSICS Reference
Define one Primary GSICS Reference instrument for each Spectral Band (VIS/NIR/TIR/UV/...) and application area (NRTC/RAC/ARC/...) ? by consensus agreement based on a set of criteria (TBD) supported by a Traceability Statement

7 Proposed Solution – Blend all Corrections to Provide Correction to Primary GSICS Reference
Define one Primary GSICS Reference Regard others as Transfer References Generate Delta Corrections* to convert to Primary Apply Delta Corrections to transfer references corrections BLEND Corrections from all available references *Delta Corrections Normally based on double differences In channel-space of monitored instruments Over extended overlap period between references

8 How? Generate RAC/NRTC/Daily Products Define Delta Correction
Using both Primary and Transfer References During whole overlap period Define Delta Correction As double difference Define Delta Correction Period Check for Jumps, Trends - Define break points if necessary Define Weights to blend references after Delta Correction Evaluate Delta Correction Uncertainty Apply Delta Corrections Blend Corrected Corrections into Primary GSICS Correction May include results from transfer refs for optional use

9 Correcting the Corrections and Blending References
Reference-1 (Primary) Monitored Instrument Reference-2 (Transfer) GSICS Correction, g1 Mon1 GSICS Correction, g2 Mon2 - Derived by GSICS Delta Correction, g1/2 21 + Corrected Correction, g2,1/2 Mon 21 Primary GSICS Correction, g0 Mon1 Mon1 Applied by User

10 Proposed Solution - Versions
If we change the Primary GSICS Reference: e.g. If a better reference becomes available in future or different applications (e.g. Archive Re-Calibration) by consensus agreement – or following selection proc. Would potentially introduce a calibration jump Generate new Version of each GSICS Product: Submit through a streamlined review process (GPPA) Revise uncertainty evaluation Revise traceability statement

11 What does this mean for the User?
Users only have to select: Monitored Instrument Spectral Band Time Period Product Version (usually most recent) (Optionally) Override default Primary GSICS Reference To select a Transfer Reference Benefits: Less to implement Less choice => Less scope for error

12 What next? Agree within GRWG general approach to :
Primary GSICS Reference Define Process to select Define Process to redefine Primary Reference Define Delta Correction & Uncertainty Define ATBD to Define Delta Correction & Uncertainty Apply Delta Correction to Corrections from Transfer References Blend Corrections from all References to generate Primary GSICS Corrections

13 Defining the Primary GSICS Reference

14 Correcting the Corrections - Defining the Delta Correction

15 Outline Delta Correction ATBD
Specific Example: GEO-LEO IR Three options to generate Delta Corrections: #1. Derive from RAC-like files #2. Derive from Daily Correction files #3. Derive from Collocations First need to define Primary GSICS Reference In practice, this will depend on the above results

16 Outline ATBD for Option #1: GEO-LEO IR - RAC
Generate intermediate RAC-like products for (Mon-Ref1) and (Mon-Ref2) Check periods over which each reference is available Define/Refine date range over which to evaluate Delta Correction Default: Whole overlap period Read coefficients from RAC (Mon-Ref1) and (Mon-Ref2) Calc Double Difference of RAC coefficients from time series of RAC(Mon-Ref2)-RAC(Mon-Ref1) Apply DD coefficients to evaluate DD bias time series for standard radiance scenes Identify any step changes in the DD bias time series – if significant: split overlap period and treat as separate reference & Go to 3 Identify any periodicity in the DD bias time series – if significant: limit date range to n periods, define uncertainty growth & Go to 3 Identify any drifts in the DD bias time series – if significant: limit date range to create period with insignificant changes, define uncertainty growth & Go to 3 If no significant changes: Define Delta Correction as mean of Double Difference of RAC coefficients and covariance Correcting covariance for oversampling in RAC time series Apply Delta Correction to (Mon-Ref2) to (Mon-Ref1/2)

17 Outline ATBD for Option #2: GEO-LEO IR - Daily
Generate intermediate Daily Correction products for (Mon-Ref1) and (Mon-Ref2) Check periods over which each reference is available Define/Refine date range over which to evaluate Delta Correction Default: Whole overlap period Read coefficients from Daily(Mon-Ref1) and (Mon-Ref2) Calc Double Difference of Daily coefficients from time series of Daily(Mon-Ref2)-Daily(Mon-Ref1) Apply DD coefficients to evaluate DD bias time series for standard radiance scenes Filter time series of DD bias time series to reject outliers Identify any step changes in the DD bias time series – if significant: split overlap period and treat as separate reference & Go to 3 Identify any periodicity in the DD bias time series – if significant: limit date range to n periods, define uncertainty growth & Go to 3 Identify any drifts in the DD bias time series – if significant: limit date range to create period with insignificant changes, define uncertainty growth & Go to 3 If no significant changes: Define Delta Correction as mean of Double Difference coefficients and covariance Apply Delta Correction to (Mon-Ref2) to generate RAC for dates when Ref1 not available

18 Outline ATBD for Option #3: GEO-LEO IR - Collocations
Generate intermediate Collocation products for (Mon-Ref1) and (Mon-Ref2) Check periods over which each reference is available Define/Refine date range over which to evaluate Delta Correction Default: Whole overlap period Recalculate RAC-like products from collocations of (Mon-Ref1) and (Mon-Ref2) Calc Double Difference of RAC coefficients from time series of RAC(Mon-Ref2)-RAC(Mon-Ref1) Apply DD coefficients to evaluate DD bias time series for standard radiance scenes Identify any step changes in the DD bias time series – if significant: split overlap period and treat as separate reference & Go to 3 Identify any periodicity in the DD bias time series – if significant: limit date range to n periods, define uncertainty growth & Go to 3 Identify any drifts in the DD bias time series – if significant: limit date range to create period with insignificant changes, define uncertainty growth & Go to 3 If no significant changes: Define Delta Correction as weighted regression from all remaining collocations coefficients and covariances Apply Delta Correction to (Mon-Ref2) to generate RAC for dates when Ref1 not available

19 Recap ATBD for Option #1: GEO-LEO IR - RAC
Generate intermediate RAC-like products for (Mon-Ref1) and (Mon-Ref2) Check periods over which each reference is available Define/Refine date range over which to evaluate Delta Correction Default: Whole overlap period Read coefficients from RAC (Mon-Ref1) and (Mon-Ref2) Calc Double Difference of RAC coefficients from time series of RAC(Mon-Ref2)-RAC(Mon-Ref1) Apply DD coefficients to evaluate DD bias time series for standard radiance scenes Identify any step changes in the DD bias time series – if significant: split overlap period and treat as separate reference & Go to 3 Identify any periodicity in the DD bias time series – if significant: limit date range to n periods, define uncertainty growth & Go to 3 Identify any drifts in the DD bias time series – if significant: limit date range to create period with insignificant changes, define uncertainty growth & Go to 3 If no significant changes: Define Delta Correction as mean of Double Difference of RAC coefficients and covariance Correcting covariance for oversampling in RAC time series Apply Delta Correction to (Mon-Ref2) to generate RAC for dates when Ref1 not available

20 ATBD for Delta Corrections
Define Delta Correction as mean of Double Difference coefficients and covariance Correcting covariance for oversampling in RAC time series Filter outliers 2- or 3-sigma filter of time series of DD std Tb biases not necessary for RAC Calculate mean and covariance of DD coefficients Scale covariance by oversampling factor where {x}t = Time Series of x <x> = Mean of Time Series of x P = Smoothing Period n = number of samples Δt= Overlap Period

21 Option #2: Daily Correction (MSG3-IASIB)-(MSG3-IASIA)

22 Option #1: Demo RAC (MSG3-IASIB)-(MSG3-IASIA)
=> Mean differences statistically consistent with Daily Results => Correction for over-sampling produces comparable uncertainties

23 Option #1: PreOp RAC (MSG3-IASIB)-(MSG3-IASIA)
More Metop overpasses => More collocations => Lower uncertainties, but consistent mean

24 Specific Example #3: GEO-LEO IR –Collocations
Generate intermediate Collocation products for (Mon-Ref1) and (Mon-Ref2) Individual collocations from each pair of instruments cannot be directly compared because they do not have the same location & time. Only their distributions can be compared statistically! E.g. Using by comparing RACs... (Option 1/2) OR Perform regression of whole overlap dataset... LREF1 LMON LREF2 LMON

25 Option #3: From Collocations
Cannot find uncertainty from time series Need to propagate from regressions Need to make assumptions about cross-covariance terms in Σ uncertainty Finding: Need to only compare matched days Otherwise, calibration trends introduce biases Reduces number of cases in my test dataset

26 Time Series of Standard Biases – Updated
Time Series of Standard Biases – Updated! Full year (Met10/SEVIRI-MetopB/IASI) & (Met10/SEVIRI-MetopA/IASI) Unsmoothed Daily Results sigma filter – not necessarily same days!

27 Recap: Option #1: Demo RAC (MSG3-IASIB)-(MSG3-IASIA)

28 Option 3: Matched Collocations (MSG3-IASIB)-(MSG3-IASIA)
Option #1: Demo RAC (MSG3-IASIB)-(MSG3-IASIA) Option 3: Matched Collocations (MSG3-IASIB)-(MSG3-IASIA) Option 3: Matched Collocations (MSG3-IASIB)-(MSG3-IASIA) Mean Biases from RAC Corrections (Option 1) closer to zero But broadly consistent with Collocations from matching days (Option 3)

29 Option #3: From Collocations - Findings
Finding: Need to only compare matched days Otherwise, calibration trends introduce biases Reduces number of cases in my test dataset Consistent results – but larger uncertainties Needs more data to be kept (all collocations) Need to make assumptions about cross-covariance terms in Σ uncertainty (Unweighted regression gives large uncertainties in shorter wavelength channels) Weighted regression: weights need increasing by a factor of ~3 to give uncertainties consistent with time series of Daily Corrections Results not robust to sub-sampling time series

30 Conclusions re: Defining Delta Corrections
Proposal (based on this analysis): Delta Corrections are calculated as: mean Double Difference of Re-Analysis Corrections Coefficients Evaluated over the whole overlap period (Option #1) Updated daily for NRTC & RAC products Delta Correction Uncertainty is calculated from: Covariance of times series of Delta Corrections Evaluated over the whole overlap period* Scaled to account for over-sampling factor (RACs are smoothed) Apply Delta Correction to all secondary references Generate GSICS Corrections to Primary Reference

31 Delta Correction Stationarity Tests
* Delta Corrections only use whole overlap period if: Test¹ shows no significant² jumps during period ¹ Suitable tests to be defined ² Significant wrt uncertainty on GSICS Correction and/or user requirements Test¹ shows no significant² cycles during period Or define for an integral number of cycles? Test¹ shows no significant² trend during period Or ... we’re in trouble...?

32 Blending Corrections from different References

33 Primary GSICS Reference Transfer Strategies
Strategy Definition Pros Cons NRTC? RAC? ARC? Single Transferable Reference Pecking Order: IASI-A, -B, -C, CrIS, AIRS,... Transition Condition V. Simple Doesn’t use all data Y Simple Blend Equal Weightings of each reference Uses all data Short periods add noise N? Y? Delayed Blend Equal Weightings + Delay F. Simple Uses most data None? (Delay) N/R Tapered Transition Tapered Weightings Smoothes steps Can’t see future! N Weighted Blend Weightings by Uncertainty Q. Simple None? Others?

34 Define Weights in Blended Reference
What matters: How good is Delta Correction relative to GSICS Correction? Evaluate Uncertainty! Uncertainty of Mean Double Difference of Overlap T/S Relative to Uncertainty of Single Difference T/S If only one reference available it gets 100% weight! If multiple references available Weight each by their total uncertainty (incl. Delta Corr.)

35 Simple Example Blended Reference
Simplest Blend: Equal Weight for each available Reference More Advanced: Tapered transition e.g. 1 year Weight based on Relative Uncertainty of Single and Double Differences

36 Recap ATBD for Option #1: GEO-LEO IR - RAC
Generate intermediate RAC-like products for (Mon-Ref1) and (Mon-Ref2) Check periods over which each reference is available Define/Refine date range over which to evaluate Delta Correction Default: Whole overlap period Read coefficients from RAC (Mon-Ref1) and (Mon-Ref2) Calc Double Difference of RAC coefficients from time series of RAC(Mon-Ref2)-RAC(Mon-Ref1) Apply DD coefficients to evaluate DD bias time series for standard radiance scenes Identify any step changes in the DD bias time series – if significant: split overlap period and treat as separate reference & Go to 3 Identify any periodicity in the DD bias time series – if significant: limit date range to n periods, define uncertainty growth & Go to 3 Identify any drifts in the DD bias time series – if significant: limit date range to create period with insignificant changes, define uncertainty growth & Go to 3 If no significant changes: Define Delta Correction as mean of Double Difference of RAC coefficients and covariance Correcting covariance for oversampling in RAC time series Apply Delta Correction to (Mon-Ref2) to generate RAC for dates when Ref1 not available How? See Next Slide...

37 How to make NRTC to Primary GSICS Reference using Delta Corrections
Check whether there is an existing Primary Reference product? If not create one For given date, check which References’ Delta Corrections and NRTCs are available for Apply Delta Correction for each available reference to each NRTC and evaluate uncertainty Combine Corrected Corrections (NRTC+Delta) for each Reference Weighted by uncertainty Write this out to Primary Reference NRTC Product

38 How to make RAC to Primary GSICS Reference using Delta Corrections
Check whether there is an existing Primary Reference product? If not create one For given date, check which References’ Delta Corrections and RACs are available for Apply Delta Correction for each available reference to each RAC and evaluate uncertainty Combine Corrected Corrections (RAC+Delta) for each Reference Weighted by uncertainty Write this out to Primary Reference RAC Product All products use common Delta Correction - mean RAC DD over whole* overlap period Advantages: Enforces a delay of 14 days before it is applied More robust than daily/NRTC Delta Corrections Simpler to implement The double differences are going to be averaged anyway! (*Or subset thereof, if jumps detected...)

39 How to make ARC to Primary GSICS Reference using Delta Corrections
Check whether there is an existing Primary Reference product? If not create one For given date, check which References’ Delta Corrections and ARCs are available for Apply Delta Correction for each available reference to each ARC and evaluate uncertainty Combine Corrected Corrections (ARC+Delta) for each Reference Weighted by uncertainty Write this out to Primary Reference ARC Product

40 Applying Delta Correction to Secondary Corrections

41 Standard Bias MSG3-IASA/IASIB/IASIB+Delta/ Primary = Avg(MSG3-IASIA,IASIB+Delta)

42 Uncertainty on blended Inter-Calibrations to Primary GSICS Reference

43 Define Primary GSICS Correction

44 Weights MSG3-IASA/IASIB+Delta

45 Uncertainty MSG3-IASA/IASIB/IASIB+Delta/ Primary = Avg(MSG3-IASIA,IASIB+Delta)
Red is usually similar to black IASIA~IASIB Blue is better than black! (More data!) Sparse data periods are filled (Jan 2014) Delta Correction improves after first months

46 Summary & Way Forward

47 Summary of Proposal Generate Delta Corrections
to convert Corrections from different references based on mean of double difference of GSICS Correction in overlap period Define one Primary GSICS Reference for each spectral band and processes to select and change Make new inter-calibration products to the Primary GSICS Reference Known as Prime GSICS Corrections blending corrections from multiple references according to uncertainties include all composite parts in single netCDF file Draft Outline ATBD & prototype products available for review

48 What Next for Delta Corrections?
Agree Delta Correction formula proposed here Define stationarity tests to characterise overlap period: Step Changes Periodic Variations – including diurnal & seasonal Long-term trends Agree how to handle these, if significant Hopefully not for IASI-A/B Final agreement on use of Primary References E.g. to Blend or Not to Blend

49 What next for Primary References?
Define process to select & migrate Primary GSICS References for each spectral band Please review draft ATBD & Prototype Products! Prepare Traceability Statement for Primary GSICS Reference, incl: Description of transfer process Comparisons with secondary references Pre-launch Characterisation Who wants to lead this?

50 What next for Blended Corrections?
Seek users’ feedback on this proposal Define netCDF format & content of GSICS Products Revise User Guides, Traceability Statement, Uncertainty Analysis Refine GSICS Procedure for Product Acceptance (GPPA) Define new structure on GSICS Data Servers Start generating products in parallel with current products Modify GSICS plotting tool to work with new products

51 Thank you Any Questions?

52 Additional Thoughts

53 Aside: How to Handle Periodic Variations
Blank out anomalous periods e.g. Midnight BB Cal Add a predictor e.g. Time of Day, Time of Year Use full number of periods to define Delta Correction To cancel out impact of variations, But uncertainties will be correspondingly larger Estimate spread empirically and inflate uncertainties to cover these

54 Aside: To Correct or Not to Correct?
Delta Corrections may not be significant i.e. Statistically or physically different from zero (no correction) E.g. For IASI-A/B in most SEVIRI channels Need to decide whether to apply insignificant Corrections Doing so can introduce unphysical variability Not doing so risks introducing jumps in time series If Correction changes to become significant Can we leave this as a decision for the users? Are you more interested in noise or long-term stability? Need to provide clear guidance

55 Aside: Combining GSICS Products
Merge Multiple Spectral Bands into one file? Treat all algorithm changes as new versions? Start with corrections only for TIR channels? Add VIS channels next, based on DCC products? Good idea for Users Difficult to implement for Developers E.g. Lots of different variables in netCDF files Could provide users with a combination of netCDFs E.g. Zip archive of pointers, ...?


Download ppt "Tim Hewison (EUMETSAT) (GRWG Chair)"

Similar presentations


Ads by Google